[BUG] AttributeError: 'PatchEmbed' object has no attribute 'dynamic_img_pad'

[Lazy-Tori]

I am using "beitv2_base_patch16_224", "beitv2_large_patch16_224".
I used it as usual, but I get an error. AttributeError: 'PatchEmbed' object has no attribute 'dynamic_img_pad'
Please fix the bug.

[rwightman]

@byunsunyoung you'll have to specify what 'using it as usual' means as it was tested and works fine for me, are you serializing the full model instead of the state dict in saved checkpoints? you cannot do that across versions reliably.

[Lazy-Tori]

I've been using the beit model since 2 weeks ago, and I've been getting errors since I restarted the server today.
I guess there was a side effect of the August 28 update.
I am using timm==0.9.5. Can you check the code that uses the "dynamic_img_pad" parameter in the beit model?

[Lazy-Tori]

For your information, I succeeded "create_model" method, and this error occurs during fine-tuning process using pytorch.

[Lazy-Tori]

Cell In[45], line 3
1 lr = 0.001
2 epochs = 30
----> 3 train_model(model, dataloaders, criterion, optimizer, num_epochs=epochs)

Cell In[44], line 37, in train_model(model, dataloaders, criterion, optimizer, num_epochs)
33 optimizer.zero_grad()
35 with torch.set_grad_enabled(phase == 'train'):
---> 37 outputs = model(inputs)
38 loss = criterion(outputs, labels)
40 _, preds = torch.max(outputs, 1)

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
1496 # If we don't have any hooks, we want to skip the rest of the logic in
1497 # this function, and just call forward.
1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1499 or _global_backward_pre_hooks or _global_backward_hooks
1500 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501 return forward_call(*args, **kwargs)
1502 # Do not call functions when jit is used
1503 full_backward_hooks, non_full_backward_hooks = [], []

File /opt/conda/lib/python3.8/site-packages/timm/models/beit.py:427, in Beit.forward(self, x)
426 def forward(self, x):
--> 427 x = self.forward_features(x)
428 x = self.forward_head(x)
429 return x

File /opt/conda/lib/python3.8/site-packages/timm/models/beit.py:404, in Beit.forward_features(self, x)
403 def forward_features(self, x):
--> 404 x = self.patch_embed(x)
405 x = torch.cat((self.cls_token.expand(x.shape[0], -1, -1), x), dim=1)
406 if self.pos_embed is not None:

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
1496 # If we don't have any hooks, we want to skip the rest of the logic in
1497 # this function, and just call forward.
1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
1499 or _global_backward_pre_hooks or _global_backward_hooks
1500 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501 return forward_call(*args, **kwargs)
1502 # Do not call functions when jit is used
1503 full_backward_hooks, non_full_backward_hooks = [], []

File /opt/conda/lib/python3.8/site-packages/timm/layers/patch_embed.py:83, in PatchEmbed.forward(self, x)
75 _assert(
76 H % self.patch_size[0] == 0,
77 f"Input height ({H}) should be divisible by patch size ({self.patch_size[0]})."
78 )
79 _assert(
80 W % self.patch_size[1] == 0,
81 f"Input width ({W}) should be divisible by patch size ({self.patch_size[1]})."
82 )
---> 83 if self.dynamic_img_pad:
84 pad_h = (self.patch_size[0] - H % self.patch_size[0]) % self.patch_size[0]
85 pad_w = (self.patch_size[1] - W % self.patch_size[1]) % self.patch_size[1]

File /opt/conda/lib/python3.8/site-packages/torch/nn/modules/module.py:1614, in Module.getattr(self, name)
1612 if name in modules:
1613 return modules[name]
-> 1614 raise AttributeError("'{}' object has no attribute '{}'".format(
1615 type(self).name, name))

AttributeError: 'PatchEmbed' object has no attribute 'dynamic_img_pad'

[rwightman]

That attribute was added in 0.9.6, but it's disabled by default and wouldn't intefere with training of beit (it's not currently used by beit). I can fine-tune that model w/o issues, see example below...

Are you sure your environment is in a good state? You don't have conflicting timm versions, this error should only happen if the model has the attributes of a previous version of timm, but is using the code for the new version somehow. Usually happens when people save their model checkpoints by doing torch.save(model) instead of torch.save(model.state_dict()) and then restore across changes in the underlying code...

python train.py --amp --amp-dtype bfloat16 --data-dir /data/standford-dogs/Images/ --num-classes 120 --model beitv2_base_patch16_224 --pretrained
Training with a single process on 1 device (cuda:0).
Loading pretrained weights from Hugging Face hub (timm/beitv2_base_patch16_224.in1k_ft_in22k_in1k)
[timm/beitv2_base_patch16_224.in1k_ft_in22k_in1k] Safe alternative available for 'pytorch_model.bin' (as 'model.safetensors'). Loading weights using safetensors.
Model beitv2_base_patch16_224 created, param count:85854264
Data processing configuration for current model + dataset:
	input_size: (3, 224, 224)
	interpolation: bicubic
	mean: (0.485, 0.456, 0.406)
	std: (0.229, 0.224, 0.225)
	crop_pct: 0.9
	crop_mode: center
Learning rate (0.05) calculated from base learning rate (0.1) and effective global batch size (128) with linear scaling.
Using native Torch AMP. Training in mixed precision.
Scheduled epochs: 300. LR stepped per epoch.
Train: 0 [   0/160 (  0%)]  Loss: 4.79 (4.79)  Time: 1.359s,   94.18/s  (1.359s,   94.18/s)  LR: 1.000e-05  Data: 0.482 (0.482)

[Lazy-Tori]

I've been using "torch.save(model)" to save the model, and there's been no problem. This is the first time there's been a problem.
I tried to restore the model using the "torch.save(model.state_dict())" as you advised, and it worked well without errors.
Thank you!

[rwightman]

@byunsunyoung I'd say that torch.save(model) IS serializing, as in it is pickling the full model class and all attributes, so when it restores to a different version of code the attributes may no longer match. torch.save(model.state_dict()) is just pickling the state dict, so only the parameters and buffers, any other python class attributes are not saved, so as long as parameter/buffer names and shapes don't change when the code does compatibility is maintained.

[rwightman]

I will move this to discussions in case anyone else runs into it, it's happened with other small layer attribute changes in the past because nuances between the two ways of saving aren't always obvious...